1#! /usr/bin/env perl
2# Copyright 2014-2020 The OpenSSL Project Authors. All Rights Reserved.
3#
4# Licensed under the OpenSSL license (the "License").  You may not use
5# this file except in compliance with the License.  You can obtain a copy
6# in the file LICENSE in the source distribution or at
7# https://www.openssl.org/source/license.html
8
9#
10# ====================================================================
11# Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
12# project. The module is, however, dual licensed under OpenSSL and
13# CRYPTOGAMS licenses depending on where you obtain it. For further
14# details see http://www.openssl.org/~appro/cryptogams/.
15# ====================================================================
16#
17# GHASH for ARMv8 Crypto Extension, 64-bit polynomial multiplication.
18#
19# June 2014
20#
21# Initial version was developed in tight cooperation with Ard
22# Biesheuvel of Linaro from bits-n-pieces from other assembly modules.
23# Just like aesv8-armx.pl this module supports both AArch32 and
24# AArch64 execution modes.
25#
26# July 2014
27#
28# Implement 2x aggregated reduction [see ghash-x86.pl for background
29# information].
30#
31# November 2017
32#
33# AArch64 register bank to "accommodate" 4x aggregated reduction and
34# improve performance by 20-70% depending on processor.
35#
36# Current performance in cycles per processed byte:
37#
38#		64-bit PMULL	32-bit PMULL	32-bit NEON(*)
39# Apple A7	0.58		0.92		5.62
40# Cortex-A53	0.85		1.01		8.39
41# Cortex-A57	0.73		1.17		7.61
42# Denver	0.51		0.65		6.02
43# Mongoose	0.65		1.10		8.06
44# Kryo		0.76		1.16		8.00
45#
46# (*)	presented for reference/comparison purposes;
47
48$flavour = shift;
49$output  = shift;
50
51$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
52( $xlate="${dir}arm-xlate.pl" and -f $xlate ) or
53( $xlate="${dir}../../perlasm/arm-xlate.pl" and -f $xlate) or
54die "can't locate arm-xlate.pl";
55
56open OUT,"| \"$^X\" $xlate $flavour $output";
57*STDOUT=*OUT;
58
59$Xi="x0";	# argument block
60$Htbl="x1";
61$inp="x2";
62$len="x3";
63
64$inc="x12";
65
66{
67my ($Xl,$Xm,$Xh,$IN)=map("q$_",(0..3));
68my ($t0,$t1,$t2,$xC2,$H,$Hhl,$H2)=map("q$_",(8..14));
69
70$code=<<___;
71#include "arm_arch.h"
72
73#if __ARM_MAX_ARCH__>=7
74.text
75___
76# $code.=".arch	armv8-a+crypto\n"	if ($flavour =~ /64/);
77$code.=<<___				if ($flavour !~ /64/);
78.fpu	neon
79.code	32
80#undef	__thumb2__
81___
82
83################################################################################
84# void gcm_init_v8(u128 Htable[16],const u64 H[2]);
85#
86# input:	128-bit H - secret parameter E(K,0^128)
87# output:	precomputed table filled with degrees of twisted H;
88#		H is twisted to handle reverse bitness of GHASH;
89#		only few of 16 slots of Htable[16] are used;
90#		data is opaque to outside world (which allows to
91#		optimize the code independently);
92#
93$code.=<<___;
94.global	gcm_init_v8
95.type	gcm_init_v8,%function
96.align	4
97gcm_init_v8:
98	vld1.64		{$t1},[x1]		@ load input H
99	vmov.i8		$xC2,#0xe1
100	vshl.i64	$xC2,$xC2,#57		@ 0xc2.0
101	vext.8		$IN,$t1,$t1,#8
102	vshr.u64	$t2,$xC2,#63
103	vdup.32		$t1,${t1}[1]
104	vext.8		$t0,$t2,$xC2,#8		@ t0=0xc2....01
105	vshr.u64	$t2,$IN,#63
106	vshr.s32	$t1,$t1,#31		@ broadcast carry bit
107	vand		$t2,$t2,$t0
108	vshl.i64	$IN,$IN,#1
109	vext.8		$t2,$t2,$t2,#8
110	vand		$t0,$t0,$t1
111	vorr		$IN,$IN,$t2		@ H<<<=1
112	veor		$H,$IN,$t0		@ twisted H
113	vst1.64		{$H},[x0],#16		@ store Htable[0]
114
115	@ calculate H^2
116	vext.8		$t0,$H,$H,#8		@ Karatsuba pre-processing
117	vpmull.p64	$Xl,$H,$H
118	veor		$t0,$t0,$H
119	vpmull2.p64	$Xh,$H,$H
120	vpmull.p64	$Xm,$t0,$t0
121
122	vext.8		$t1,$Xl,$Xh,#8		@ Karatsuba post-processing
123	veor		$t2,$Xl,$Xh
124	veor		$Xm,$Xm,$t1
125	veor		$Xm,$Xm,$t2
126	vpmull.p64	$t2,$Xl,$xC2		@ 1st phase
127
128	vmov		$Xh#lo,$Xm#hi		@ Xh|Xm - 256-bit result
129	vmov		$Xm#hi,$Xl#lo		@ Xm is rotated Xl
130	veor		$Xl,$Xm,$t2
131
132	vext.8		$t2,$Xl,$Xl,#8		@ 2nd phase
133	vpmull.p64	$Xl,$Xl,$xC2
134	veor		$t2,$t2,$Xh
135	veor		$H2,$Xl,$t2
136
137	vext.8		$t1,$H2,$H2,#8		@ Karatsuba pre-processing
138	veor		$t1,$t1,$H2
139	vext.8		$Hhl,$t0,$t1,#8		@ pack Karatsuba pre-processed
140	vst1.64		{$Hhl-$H2},[x0],#32	@ store Htable[1..2]
141___
142if ($flavour =~ /64/) {
143my ($t3,$Yl,$Ym,$Yh) = map("q$_",(4..7));
144
145$code.=<<___;
146	@ calculate H^3 and H^4
147	vpmull.p64	$Xl,$H, $H2
148	 vpmull.p64	$Yl,$H2,$H2
149	vpmull2.p64	$Xh,$H, $H2
150	 vpmull2.p64	$Yh,$H2,$H2
151	vpmull.p64	$Xm,$t0,$t1
152	 vpmull.p64	$Ym,$t1,$t1
153
154	vext.8		$t0,$Xl,$Xh,#8		@ Karatsuba post-processing
155	 vext.8		$t1,$Yl,$Yh,#8
156	veor		$t2,$Xl,$Xh
157	veor		$Xm,$Xm,$t0
158	 veor		$t3,$Yl,$Yh
159	 veor		$Ym,$Ym,$t1
160	veor		$Xm,$Xm,$t2
161	vpmull.p64	$t2,$Xl,$xC2		@ 1st phase
162	 veor		$Ym,$Ym,$t3
163	 vpmull.p64	$t3,$Yl,$xC2
164
165	vmov		$Xh#lo,$Xm#hi		@ Xh|Xm - 256-bit result
166	 vmov		$Yh#lo,$Ym#hi
167	vmov		$Xm#hi,$Xl#lo		@ Xm is rotated Xl
168	 vmov		$Ym#hi,$Yl#lo
169	veor		$Xl,$Xm,$t2
170	 veor		$Yl,$Ym,$t3
171
172	vext.8		$t2,$Xl,$Xl,#8		@ 2nd phase
173	 vext.8		$t3,$Yl,$Yl,#8
174	vpmull.p64	$Xl,$Xl,$xC2
175	 vpmull.p64	$Yl,$Yl,$xC2
176	veor		$t2,$t2,$Xh
177	 veor		$t3,$t3,$Yh
178	veor		$H, $Xl,$t2		@ H^3
179	 veor		$H2,$Yl,$t3		@ H^4
180
181	vext.8		$t0,$H, $H,#8		@ Karatsuba pre-processing
182	 vext.8		$t1,$H2,$H2,#8
183	veor		$t0,$t0,$H
184	 veor		$t1,$t1,$H2
185	vext.8		$Hhl,$t0,$t1,#8		@ pack Karatsuba pre-processed
186	vst1.64		{$H-$H2},[x0]		@ store Htable[3..5]
187___
188}
189$code.=<<___;
190	ret
191.size	gcm_init_v8,.-gcm_init_v8
192___
193################################################################################
194# void gcm_gmult_v8(u64 Xi[2],const u128 Htable[16]);
195#
196# input:	Xi - current hash value;
197#		Htable - table precomputed in gcm_init_v8;
198# output:	Xi - next hash value Xi;
199#
200$code.=<<___;
201.global	gcm_gmult_v8
202.type	gcm_gmult_v8,%function
203.align	4
204gcm_gmult_v8:
205	vld1.64		{$t1},[$Xi]		@ load Xi
206	vmov.i8		$xC2,#0xe1
207	vld1.64		{$H-$Hhl},[$Htbl]	@ load twisted H, ...
208	vshl.u64	$xC2,$xC2,#57
209#ifndef __ARMEB__
210	vrev64.8	$t1,$t1
211#endif
212	vext.8		$IN,$t1,$t1,#8
213
214	vpmull.p64	$Xl,$H,$IN		@ H.lo·Xi.lo
215	veor		$t1,$t1,$IN		@ Karatsuba pre-processing
216	vpmull2.p64	$Xh,$H,$IN		@ H.hi·Xi.hi
217	vpmull.p64	$Xm,$Hhl,$t1		@ (H.lo+H.hi)·(Xi.lo+Xi.hi)
218
219	vext.8		$t1,$Xl,$Xh,#8		@ Karatsuba post-processing
220	veor		$t2,$Xl,$Xh
221	veor		$Xm,$Xm,$t1
222	veor		$Xm,$Xm,$t2
223	vpmull.p64	$t2,$Xl,$xC2		@ 1st phase of reduction
224
225	vmov		$Xh#lo,$Xm#hi		@ Xh|Xm - 256-bit result
226	vmov		$Xm#hi,$Xl#lo		@ Xm is rotated Xl
227	veor		$Xl,$Xm,$t2
228
229	vext.8		$t2,$Xl,$Xl,#8		@ 2nd phase of reduction
230	vpmull.p64	$Xl,$Xl,$xC2
231	veor		$t2,$t2,$Xh
232	veor		$Xl,$Xl,$t2
233
234#ifndef __ARMEB__
235	vrev64.8	$Xl,$Xl
236#endif
237	vext.8		$Xl,$Xl,$Xl,#8
238	vst1.64		{$Xl},[$Xi]		@ write out Xi
239
240	ret
241.size	gcm_gmult_v8,.-gcm_gmult_v8
242___
243################################################################################
244# void gcm_ghash_v8(u64 Xi[2],const u128 Htable[16],const u8 *inp,size_t len);
245#
246# input:	table precomputed in gcm_init_v8;
247#		current hash value Xi;
248#		pointer to input data;
249#		length of input data in bytes, but divisible by block size;
250# output:	next hash value Xi;
251#
252$code.=<<___;
253.global	gcm_ghash_v8
254.type	gcm_ghash_v8,%function
255.align	4
256gcm_ghash_v8:
257___
258$code.=<<___	if ($flavour =~ /64/);
259	cmp		$len,#64
260	b.hs		.Lgcm_ghash_v8_4x
261___
262$code.=<<___		if ($flavour !~ /64/);
263	vstmdb		sp!,{d8-d15}		@ 32-bit ABI says so
264___
265$code.=<<___;
266	vld1.64		{$Xl},[$Xi]		@ load [rotated] Xi
267						@ "[rotated]" means that
268						@ loaded value would have
269						@ to be rotated in order to
270						@ make it appear as in
271						@ algorithm specification
272	subs		$len,$len,#32		@ see if $len is 32 or larger
273	mov		$inc,#16		@ $inc is used as post-
274						@ increment for input pointer;
275						@ as loop is modulo-scheduled
276						@ $inc is zeroed just in time
277						@ to preclude overstepping
278						@ inp[len], which means that
279						@ last block[s] are actually
280						@ loaded twice, but last
281						@ copy is not processed
282	vld1.64		{$H-$Hhl},[$Htbl],#32	@ load twisted H, ..., H^2
283	vmov.i8		$xC2,#0xe1
284	vld1.64		{$H2},[$Htbl]
285	cclr		$inc,eq			@ is it time to zero $inc?
286	vext.8		$Xl,$Xl,$Xl,#8		@ rotate Xi
287	vld1.64		{$t0},[$inp],#16	@ load [rotated] I[0]
288	vshl.u64	$xC2,$xC2,#57		@ compose 0xc2.0 constant
289#ifndef __ARMEB__
290	vrev64.8	$t0,$t0
291	vrev64.8	$Xl,$Xl
292#endif
293	vext.8		$IN,$t0,$t0,#8		@ rotate I[0]
294	b.lo		.Lodd_tail_v8		@ $len was less than 32
295___
296{ my ($Xln,$Xmn,$Xhn,$In) = map("q$_",(4..7));
297	#######
298	# Xi+2 =[H*(Ii+1 + Xi+1)] mod P =
299	#	[(H*Ii+1) + (H*Xi+1)] mod P =
300	#	[(H*Ii+1) + H^2*(Ii+Xi)] mod P
301	#
302$code.=<<___;
303	vld1.64		{$t1},[$inp],$inc	@ load [rotated] I[1]
304#ifndef __ARMEB__
305	vrev64.8	$t1,$t1
306#endif
307	vext.8		$In,$t1,$t1,#8
308	veor		$IN,$IN,$Xl		@ I[i]^=Xi
309	vpmull.p64	$Xln,$H,$In		@ H·Ii+1
310	veor		$t1,$t1,$In		@ Karatsuba pre-processing
311	vpmull2.p64	$Xhn,$H,$In
312	b		.Loop_mod2x_v8
313
314.align	4
315.Loop_mod2x_v8:
316	vext.8		$t2,$IN,$IN,#8
317	subs		$len,$len,#32		@ is there more data?
318	vpmull.p64	$Xl,$H2,$IN		@ H^2.lo·Xi.lo
319	cclr		$inc,lo			@ is it time to zero $inc?
320
321	 vpmull.p64	$Xmn,$Hhl,$t1
322	veor		$t2,$t2,$IN		@ Karatsuba pre-processing
323	vpmull2.p64	$Xh,$H2,$IN		@ H^2.hi·Xi.hi
324	veor		$Xl,$Xl,$Xln		@ accumulate
325	vpmull2.p64	$Xm,$Hhl,$t2		@ (H^2.lo+H^2.hi)·(Xi.lo+Xi.hi)
326	 vld1.64	{$t0},[$inp],$inc	@ load [rotated] I[i+2]
327
328	veor		$Xh,$Xh,$Xhn
329	 cclr		$inc,eq			@ is it time to zero $inc?
330	veor		$Xm,$Xm,$Xmn
331
332	vext.8		$t1,$Xl,$Xh,#8		@ Karatsuba post-processing
333	veor		$t2,$Xl,$Xh
334	veor		$Xm,$Xm,$t1
335	 vld1.64	{$t1},[$inp],$inc	@ load [rotated] I[i+3]
336#ifndef __ARMEB__
337	 vrev64.8	$t0,$t0
338#endif
339	veor		$Xm,$Xm,$t2
340	vpmull.p64	$t2,$Xl,$xC2		@ 1st phase of reduction
341
342#ifndef __ARMEB__
343	 vrev64.8	$t1,$t1
344#endif
345	vmov		$Xh#lo,$Xm#hi		@ Xh|Xm - 256-bit result
346	vmov		$Xm#hi,$Xl#lo		@ Xm is rotated Xl
347	 vext.8		$In,$t1,$t1,#8
348	 vext.8		$IN,$t0,$t0,#8
349	veor		$Xl,$Xm,$t2
350	 vpmull.p64	$Xln,$H,$In		@ H·Ii+1
351	veor		$IN,$IN,$Xh		@ accumulate $IN early
352
353	vext.8		$t2,$Xl,$Xl,#8		@ 2nd phase of reduction
354	vpmull.p64	$Xl,$Xl,$xC2
355	veor		$IN,$IN,$t2
356	 veor		$t1,$t1,$In		@ Karatsuba pre-processing
357	veor		$IN,$IN,$Xl
358	 vpmull2.p64	$Xhn,$H,$In
359	b.hs		.Loop_mod2x_v8		@ there was at least 32 more bytes
360
361	veor		$Xh,$Xh,$t2
362	vext.8		$IN,$t0,$t0,#8		@ re-construct $IN
363	adds		$len,$len,#32		@ re-construct $len
364	veor		$Xl,$Xl,$Xh		@ re-construct $Xl
365	b.eq		.Ldone_v8		@ is $len zero?
366___
367}
368$code.=<<___;
369.Lodd_tail_v8:
370	vext.8		$t2,$Xl,$Xl,#8
371	veor		$IN,$IN,$Xl		@ inp^=Xi
372	veor		$t1,$t0,$t2		@ $t1 is rotated inp^Xi
373
374	vpmull.p64	$Xl,$H,$IN		@ H.lo·Xi.lo
375	veor		$t1,$t1,$IN		@ Karatsuba pre-processing
376	vpmull2.p64	$Xh,$H,$IN		@ H.hi·Xi.hi
377	vpmull.p64	$Xm,$Hhl,$t1		@ (H.lo+H.hi)·(Xi.lo+Xi.hi)
378
379	vext.8		$t1,$Xl,$Xh,#8		@ Karatsuba post-processing
380	veor		$t2,$Xl,$Xh
381	veor		$Xm,$Xm,$t1
382	veor		$Xm,$Xm,$t2
383	vpmull.p64	$t2,$Xl,$xC2		@ 1st phase of reduction
384
385	vmov		$Xh#lo,$Xm#hi		@ Xh|Xm - 256-bit result
386	vmov		$Xm#hi,$Xl#lo		@ Xm is rotated Xl
387	veor		$Xl,$Xm,$t2
388
389	vext.8		$t2,$Xl,$Xl,#8		@ 2nd phase of reduction
390	vpmull.p64	$Xl,$Xl,$xC2
391	veor		$t2,$t2,$Xh
392	veor		$Xl,$Xl,$t2
393
394.Ldone_v8:
395#ifndef __ARMEB__
396	vrev64.8	$Xl,$Xl
397#endif
398	vext.8		$Xl,$Xl,$Xl,#8
399	vst1.64		{$Xl},[$Xi]		@ write out Xi
400
401___
402$code.=<<___		if ($flavour !~ /64/);
403	vldmia		sp!,{d8-d15}		@ 32-bit ABI says so
404___
405$code.=<<___;
406	ret
407.size	gcm_ghash_v8,.-gcm_ghash_v8
408___
409
410if ($flavour =~ /64/) {				# 4x subroutine
411my ($I0,$j1,$j2,$j3,
412    $I1,$I2,$I3,$H3,$H34,$H4,$Yl,$Ym,$Yh) = map("q$_",(4..7,15..23));
413
414$code.=<<___;
415.type	gcm_ghash_v8_4x,%function
416.align	4
417gcm_ghash_v8_4x:
418.Lgcm_ghash_v8_4x:
419	vld1.64		{$Xl},[$Xi]		@ load [rotated] Xi
420	vld1.64		{$H-$H2},[$Htbl],#48	@ load twisted H, ..., H^2
421	vmov.i8		$xC2,#0xe1
422	vld1.64		{$H3-$H4},[$Htbl]	@ load twisted H^3, ..., H^4
423	vshl.u64	$xC2,$xC2,#57		@ compose 0xc2.0 constant
424
425	vld1.64		{$I0-$j3},[$inp],#64
426#ifndef __ARMEB__
427	vrev64.8	$Xl,$Xl
428	vrev64.8	$j1,$j1
429	vrev64.8	$j2,$j2
430	vrev64.8	$j3,$j3
431	vrev64.8	$I0,$I0
432#endif
433	vext.8		$I3,$j3,$j3,#8
434	vext.8		$I2,$j2,$j2,#8
435	vext.8		$I1,$j1,$j1,#8
436
437	vpmull.p64	$Yl,$H,$I3		@ H·Ii+3
438	veor		$j3,$j3,$I3
439	vpmull2.p64	$Yh,$H,$I3
440	vpmull.p64	$Ym,$Hhl,$j3
441
442	vpmull.p64	$t0,$H2,$I2		@ H^2·Ii+2
443	veor		$j2,$j2,$I2
444	vpmull2.p64	$I2,$H2,$I2
445	vpmull2.p64	$j2,$Hhl,$j2
446
447	veor		$Yl,$Yl,$t0
448	veor		$Yh,$Yh,$I2
449	veor		$Ym,$Ym,$j2
450
451	vpmull.p64	$j3,$H3,$I1		@ H^3·Ii+1
452	veor		$j1,$j1,$I1
453	vpmull2.p64	$I1,$H3,$I1
454	vpmull.p64	$j1,$H34,$j1
455
456	veor		$Yl,$Yl,$j3
457	veor		$Yh,$Yh,$I1
458	veor		$Ym,$Ym,$j1
459
460	subs		$len,$len,#128
461	b.lo		.Ltail4x
462
463	b		.Loop4x
464
465.align	4
466.Loop4x:
467	veor		$t0,$I0,$Xl
468	 vld1.64	{$I0-$j3},[$inp],#64
469	vext.8		$IN,$t0,$t0,#8
470#ifndef __ARMEB__
471	 vrev64.8	$j1,$j1
472	 vrev64.8	$j2,$j2
473	 vrev64.8	$j3,$j3
474	 vrev64.8	$I0,$I0
475#endif
476
477	vpmull.p64	$Xl,$H4,$IN		@ H^4·(Xi+Ii)
478	veor		$t0,$t0,$IN
479	vpmull2.p64	$Xh,$H4,$IN
480	 vext.8		$I3,$j3,$j3,#8
481	vpmull2.p64	$Xm,$H34,$t0
482
483	veor		$Xl,$Xl,$Yl
484	veor		$Xh,$Xh,$Yh
485	 vext.8		$I2,$j2,$j2,#8
486	veor		$Xm,$Xm,$Ym
487	 vext.8		$I1,$j1,$j1,#8
488
489	vext.8		$t1,$Xl,$Xh,#8		@ Karatsuba post-processing
490	veor		$t2,$Xl,$Xh
491	 vpmull.p64	$Yl,$H,$I3		@ H·Ii+3
492	 veor		$j3,$j3,$I3
493	veor		$Xm,$Xm,$t1
494	 vpmull2.p64	$Yh,$H,$I3
495	veor		$Xm,$Xm,$t2
496	 vpmull.p64	$Ym,$Hhl,$j3
497
498	vpmull.p64	$t2,$Xl,$xC2		@ 1st phase of reduction
499	vmov		$Xh#lo,$Xm#hi		@ Xh|Xm - 256-bit result
500	vmov		$Xm#hi,$Xl#lo		@ Xm is rotated Xl
501	 vpmull.p64	$t0,$H2,$I2		@ H^2·Ii+2
502	 veor		$j2,$j2,$I2
503	 vpmull2.p64	$I2,$H2,$I2
504	veor		$Xl,$Xm,$t2
505	 vpmull2.p64	$j2,$Hhl,$j2
506
507	 veor		$Yl,$Yl,$t0
508	 veor		$Yh,$Yh,$I2
509	 veor		$Ym,$Ym,$j2
510
511	vext.8		$t2,$Xl,$Xl,#8		@ 2nd phase of reduction
512	vpmull.p64	$Xl,$Xl,$xC2
513	 vpmull.p64	$j3,$H3,$I1		@ H^3·Ii+1
514	 veor		$j1,$j1,$I1
515	veor		$t2,$t2,$Xh
516	 vpmull2.p64	$I1,$H3,$I1
517	 vpmull.p64	$j1,$H34,$j1
518
519	veor		$Xl,$Xl,$t2
520	 veor		$Yl,$Yl,$j3
521	 veor		$Yh,$Yh,$I1
522	vext.8		$Xl,$Xl,$Xl,#8
523	 veor		$Ym,$Ym,$j1
524
525	subs		$len,$len,#64
526	b.hs		.Loop4x
527
528.Ltail4x:
529	veor		$t0,$I0,$Xl
530	vext.8		$IN,$t0,$t0,#8
531
532	vpmull.p64	$Xl,$H4,$IN		@ H^4·(Xi+Ii)
533	veor		$t0,$t0,$IN
534	vpmull2.p64	$Xh,$H4,$IN
535	vpmull2.p64	$Xm,$H34,$t0
536
537	veor		$Xl,$Xl,$Yl
538	veor		$Xh,$Xh,$Yh
539	veor		$Xm,$Xm,$Ym
540
541	adds		$len,$len,#64
542	b.eq		.Ldone4x
543
544	cmp		$len,#32
545	b.lo		.Lone
546	b.eq		.Ltwo
547.Lthree:
548	vext.8		$t1,$Xl,$Xh,#8		@ Karatsuba post-processing
549	veor		$t2,$Xl,$Xh
550	veor		$Xm,$Xm,$t1
551	 vld1.64	{$I0-$j2},[$inp]
552	veor		$Xm,$Xm,$t2
553#ifndef	__ARMEB__
554	 vrev64.8	$j1,$j1
555	 vrev64.8	$j2,$j2
556	 vrev64.8	$I0,$I0
557#endif
558
559	vpmull.p64	$t2,$Xl,$xC2		@ 1st phase of reduction
560	vmov		$Xh#lo,$Xm#hi		@ Xh|Xm - 256-bit result
561	vmov		$Xm#hi,$Xl#lo		@ Xm is rotated Xl
562	 vext.8		$I2,$j2,$j2,#8
563	 vext.8		$I1,$j1,$j1,#8
564	veor		$Xl,$Xm,$t2
565
566	 vpmull.p64	$Yl,$H,$I2		@ H·Ii+2
567	 veor		$j2,$j2,$I2
568
569	vext.8		$t2,$Xl,$Xl,#8		@ 2nd phase of reduction
570	vpmull.p64	$Xl,$Xl,$xC2
571	veor		$t2,$t2,$Xh
572	 vpmull2.p64	$Yh,$H,$I2
573	 vpmull.p64	$Ym,$Hhl,$j2
574	veor		$Xl,$Xl,$t2
575	 vpmull.p64	$j3,$H2,$I1		@ H^2·Ii+1
576	 veor		$j1,$j1,$I1
577	vext.8		$Xl,$Xl,$Xl,#8
578
579	 vpmull2.p64	$I1,$H2,$I1
580	veor		$t0,$I0,$Xl
581	 vpmull2.p64	$j1,$Hhl,$j1
582	vext.8		$IN,$t0,$t0,#8
583
584	 veor		$Yl,$Yl,$j3
585	 veor		$Yh,$Yh,$I1
586	 veor		$Ym,$Ym,$j1
587
588	vpmull.p64	$Xl,$H3,$IN		@ H^3·(Xi+Ii)
589	veor		$t0,$t0,$IN
590	vpmull2.p64	$Xh,$H3,$IN
591	vpmull.p64	$Xm,$H34,$t0
592
593	veor		$Xl,$Xl,$Yl
594	veor		$Xh,$Xh,$Yh
595	veor		$Xm,$Xm,$Ym
596	b		.Ldone4x
597
598.align	4
599.Ltwo:
600	vext.8		$t1,$Xl,$Xh,#8		@ Karatsuba post-processing
601	veor		$t2,$Xl,$Xh
602	veor		$Xm,$Xm,$t1
603	 vld1.64	{$I0-$j1},[$inp]
604	veor		$Xm,$Xm,$t2
605#ifndef	__ARMEB__
606	 vrev64.8	$j1,$j1
607	 vrev64.8	$I0,$I0
608#endif
609
610	vpmull.p64	$t2,$Xl,$xC2		@ 1st phase of reduction
611	vmov		$Xh#lo,$Xm#hi		@ Xh|Xm - 256-bit result
612	vmov		$Xm#hi,$Xl#lo		@ Xm is rotated Xl
613	 vext.8		$I1,$j1,$j1,#8
614	veor		$Xl,$Xm,$t2
615
616	vext.8		$t2,$Xl,$Xl,#8		@ 2nd phase of reduction
617	vpmull.p64	$Xl,$Xl,$xC2
618	veor		$t2,$t2,$Xh
619	veor		$Xl,$Xl,$t2
620	vext.8		$Xl,$Xl,$Xl,#8
621
622	 vpmull.p64	$Yl,$H,$I1		@ H·Ii+1
623	 veor		$j1,$j1,$I1
624
625	veor		$t0,$I0,$Xl
626	vext.8		$IN,$t0,$t0,#8
627
628	 vpmull2.p64	$Yh,$H,$I1
629	 vpmull.p64	$Ym,$Hhl,$j1
630
631	vpmull.p64	$Xl,$H2,$IN		@ H^2·(Xi+Ii)
632	veor		$t0,$t0,$IN
633	vpmull2.p64	$Xh,$H2,$IN
634	vpmull2.p64	$Xm,$Hhl,$t0
635
636	veor		$Xl,$Xl,$Yl
637	veor		$Xh,$Xh,$Yh
638	veor		$Xm,$Xm,$Ym
639	b		.Ldone4x
640
641.align	4
642.Lone:
643	vext.8		$t1,$Xl,$Xh,#8		@ Karatsuba post-processing
644	veor		$t2,$Xl,$Xh
645	veor		$Xm,$Xm,$t1
646	 vld1.64	{$I0},[$inp]
647	veor		$Xm,$Xm,$t2
648#ifndef	__ARMEB__
649	 vrev64.8	$I0,$I0
650#endif
651
652	vpmull.p64	$t2,$Xl,$xC2		@ 1st phase of reduction
653	vmov		$Xh#lo,$Xm#hi		@ Xh|Xm - 256-bit result
654	vmov		$Xm#hi,$Xl#lo		@ Xm is rotated Xl
655	veor		$Xl,$Xm,$t2
656
657	vext.8		$t2,$Xl,$Xl,#8		@ 2nd phase of reduction
658	vpmull.p64	$Xl,$Xl,$xC2
659	veor		$t2,$t2,$Xh
660	veor		$Xl,$Xl,$t2
661	vext.8		$Xl,$Xl,$Xl,#8
662
663	veor		$t0,$I0,$Xl
664	vext.8		$IN,$t0,$t0,#8
665
666	vpmull.p64	$Xl,$H,$IN
667	veor		$t0,$t0,$IN
668	vpmull2.p64	$Xh,$H,$IN
669	vpmull.p64	$Xm,$Hhl,$t0
670
671.Ldone4x:
672	vext.8		$t1,$Xl,$Xh,#8		@ Karatsuba post-processing
673	veor		$t2,$Xl,$Xh
674	veor		$Xm,$Xm,$t1
675	veor		$Xm,$Xm,$t2
676
677	vpmull.p64	$t2,$Xl,$xC2		@ 1st phase of reduction
678	vmov		$Xh#lo,$Xm#hi		@ Xh|Xm - 256-bit result
679	vmov		$Xm#hi,$Xl#lo		@ Xm is rotated Xl
680	veor		$Xl,$Xm,$t2
681
682	vext.8		$t2,$Xl,$Xl,#8		@ 2nd phase of reduction
683	vpmull.p64	$Xl,$Xl,$xC2
684	veor		$t2,$t2,$Xh
685	veor		$Xl,$Xl,$t2
686	vext.8		$Xl,$Xl,$Xl,#8
687
688#ifndef __ARMEB__
689	vrev64.8	$Xl,$Xl
690#endif
691	vst1.64		{$Xl},[$Xi]		@ write out Xi
692
693	ret
694.size	gcm_ghash_v8_4x,.-gcm_ghash_v8_4x
695___
696
697}
698}
699
700$code.=<<___;
701.asciz  "GHASH for ARMv8, CRYPTOGAMS by <appro\@openssl.org>"
702.align  2
703#endif
704___
705
706if ($flavour =~ /64/) {			######## 64-bit code
707    sub unvmov {
708	my $arg=shift;
709
710	$arg =~ m/q([0-9]+)#(lo|hi),\s*q([0-9]+)#(lo|hi)/o &&
711	sprintf	"ins	v%d.d[%d],v%d.d[%d]",$1<8?$1:$1+8,($2 eq "lo")?0:1,
712					     $3<8?$3:$3+8,($4 eq "lo")?0:1;
713    }
714    foreach(split("\n",$code)) {
715	s/cclr\s+([wx])([^,]+),\s*([a-z]+)/csel	$1$2,$1zr,$1$2,$3/o	or
716	s/vmov\.i8/movi/o		or	# fix up legacy mnemonics
717	s/vmov\s+(.*)/unvmov($1)/geo	or
718	s/vext\.8/ext/o			or
719	s/vshr\.s/sshr\.s/o		or
720	s/vshr/ushr/o			or
721	s/^(\s+)v/$1/o			or	# strip off v prefix
722	s/\bbx\s+lr\b/ret/o;
723
724	s/\bq([0-9]+)\b/"v".($1<8?$1:$1+8).".16b"/geo;	# old->new registers
725	s/@\s/\/\//o;				# old->new style commentary
726
727	# fix up remaining legacy suffixes
728	s/\.[ui]?8(\s)/$1/o;
729	s/\.[uis]?32//o and s/\.16b/\.4s/go;
730	m/\.p64/o and s/\.16b/\.1q/o;		# 1st pmull argument
731	m/l\.p64/o and s/\.16b/\.1d/go;		# 2nd and 3rd pmull arguments
732	s/\.[uisp]?64//o and s/\.16b/\.2d/go;
733	s/\.[42]([sd])\[([0-3])\]/\.$1\[$2\]/o;
734
735	print $_,"\n";
736    }
737} else {				######## 32-bit code
738    sub unvdup32 {
739	my $arg=shift;
740
741	$arg =~ m/q([0-9]+),\s*q([0-9]+)\[([0-3])\]/o &&
742	sprintf	"vdup.32	q%d,d%d[%d]",$1,2*$2+($3>>1),$3&1;
743    }
744    sub unvpmullp64 {
745	my ($mnemonic,$arg)=@_;
746
747	if ($arg =~ m/q([0-9]+),\s*q([0-9]+),\s*q([0-9]+)/o) {
748	    my $word = 0xf2a00e00|(($1&7)<<13)|(($1&8)<<19)
749				 |(($2&7)<<17)|(($2&8)<<4)
750				 |(($3&7)<<1) |(($3&8)<<2);
751	    $word |= 0x00010001	 if ($mnemonic =~ "2");
752	    # since ARMv7 instructions are always encoded little-endian.
753	    # correct solution is to use .inst directive, but older
754	    # assemblers don't implement it:-(
755	    sprintf ".byte\t0x%02x,0x%02x,0x%02x,0x%02x\t@ %s %s",
756			$word&0xff,($word>>8)&0xff,
757			($word>>16)&0xff,($word>>24)&0xff,
758			$mnemonic,$arg;
759	}
760    }
761
762    foreach(split("\n",$code)) {
763	s/\b[wx]([0-9]+)\b/r$1/go;		# new->old registers
764	s/\bv([0-9])\.[12468]+[bsd]\b/q$1/go;	# new->old registers
765	s/\/\/\s?/@ /o;				# new->old style commentary
766
767	# fix up remaining new-style suffixes
768	s/\],#[0-9]+/]!/o;
769
770	s/cclr\s+([^,]+),\s*([a-z]+)/mov$2	$1,#0/o			or
771	s/vdup\.32\s+(.*)/unvdup32($1)/geo				or
772	s/v?(pmull2?)\.p64\s+(.*)/unvpmullp64($1,$2)/geo		or
773	s/\bq([0-9]+)#(lo|hi)/sprintf "d%d",2*$1+($2 eq "hi")/geo	or
774	s/^(\s+)b\./$1b/o						or
775	s/^(\s+)ret/$1bx\tlr/o;
776
777	print $_,"\n";
778    }
779}
780
781close STDOUT or die "error closing STDOUT: $!"; # enforce flush
782